Square Packing Atlas
how a floor gets proved

Proving that no better packing exists

Finding a good packing is a search. Proving that nothing smaller can work is a different job: you have to rule out every arrangement at once, including ones nobody has thought of. Almost every proof in this subject uses one trick, due to Frits Göbel (1979): unavoidable points.

The trick

Suppose you want to show that n unit squares can't fit in a square box of side s. Scatter some points in the box so that every possible position of a unit square — anywhere, at any angle — covers at least one of them. If you managed it with fewer than n points, you're done: n squares with no overlap would each need their own point, and there aren't enough to go round.

The whole art is choosing the points. Near a wall a unit square can't tilt much without poking out, so a row of points about 0.9 from each wall catches everything there. In the middle, a triangle of points with sides ≤ 1 can't be dodged (a unit square whose centre is in the triangle must cover a corner of it). Stitching such lemmas together over the whole box is the proof.

A sharper version asks each point to be hit not once but k times: if every unit square covers at least k of m points, then at most ⌊m/k⌋ squares fit. Below is a real one.

Try it: 81 points, every square catches 7

In a box of side 35/9 ≈ 3.889 there are 81 points such that every unit square inside the box, at every position and angle, contains at least 7 of them. Twelve squares would need 12 × 7 = 84. So twelve unit squares fit in no box smaller than this: s(12) ≥ 35/9.

Drag the square around and rotate it. The points it covers light up and the count is shown. Try to get it below 7 — you can't, but the places where it is exactly 7 tell you what the points are for.

points covered
—
exactly 789–1011+

The map samples 120 × 120 positions; the real proof checks the whole continuum of positions and angles with exact integer arithmetic (an open-source verifier). The Lean proof assistant checks it all again, the covering property included, inside its kernel: s(12) ≥ 35/9 is a Lean theorem with no hypothesis. Points exactly on the square's edge count as covered. That convention is what lets a packing in any smaller box be scaled up into this one and refuted, which is why the conclusion covers every box smaller than 35/9. The skeleton is the same in every such set found: a # of lines one unit from each wall, and a small ring at the centre.

Why twelve is still open

Everyone believes s(12) = 4: the plain 4 × 4 grid with four squares missing. The best proven floor is now 3.9715: our own certificate (3.9686, computer-checked and kernel-checked in Lean, not refereed), with its points dilated and re-weighted by squarepacker (replayed by jlevy/squares). It uses a weighted version of the same argument, where a linear program chooses fractional weights for the points, an idea from Sam Burns's and Gustavo Massaccesi's 2026 work on n = 17 (Sources §5). Weighted points alone cannot finish the job: at box side 4 no weighting of points totals less than 12.27 (proved exactly). Closing the last 0.03 needs something more, such as the case analysis used to prove s(13) = 4, next.

Walkthrough: how Bentz ruled out thirteen

Bentz works in [0,4]² with boxes, each the inside of a square of side a little more than 1. If 13 boxes never fit, then s(13) = 4. He starts from 16 unavoidable points (in the figure). Thirteen boxes against sixteen points leave three points of slack. That is enough to force at least two boxes to sit alone with a corner point, where a replacement lemma pins them down further. The argument then splits on whether those two boxes share a corner, and the split has six leaves.

Click through the leaves and watch the number in the tree: every branch ends at 12. The bookkeeping is "free points + boxes already placed", and a contradiction needs that total to be under 13. Each configuration Bentz builds is, in effect, a certificate worth exactly 12 — so it kills a thirteenth box and nothing more. For 12 squares every leaf would have to come out at 11 instead, and only one of the six does. The loss starts at the very first step: with 12 boxes and 16 points the slack is 4, all eight corner points can pair up in four boxes, and there is no lonely box for the replacement lemma to work on. That single missing unit is why the same proof says nothing about s(12).

Thirteen again, with no cases

The six leaves above are the price of plain points. Give the points fractional weights, chosen by a linear program, and the case tree disappears: 3,621 weighted points in [0,4]², total 12.956, such that every closed unit square in the box captures weight at least 1. Thirteen squares would need 13. A point on a square's edge counts, which lets neighbouring tiles share weight. The margin is then zero, so the check is an exact subdivision of every position and angle, done by two separately written checkers and by Lean's kernel. The theorem is Bentz's; only the proof is new: Thirteen Squares.

What the method has proved since

Twelve and eleven squares. Weighted points in a box a little smaller than 4 give s(12) ≥ 15680/3951 = 3.9686 (August 2026). Re-solving the weights of the same points exactly gives 15680000/3949423 = 3.9702 (jlevy/squares, 3 October 2026), and dilating the points and re-weighting them on a finer angle net gives 7943/2000 = 3.9715 (squarepacker, 5 October), the best floor for twelve squares we know of. Ours, 3.9686, is the one kernel-checked in Lean. The same method gave s(11) ≥ 3.8143, since superseded by jlevy's 3.827, Kleddamag's 3.875 and, on 2026-09-29, Queuingtheorydotcom's proof that Trump's 1979 packing is optimal.

Thirteen squares. The case-free proof above, the first of our covers of a box of side exactly k, where there is no margin to spend. It is kernel-checked in Lean; an earlier Lean proof of s(13) = 4, following Bentz's argument, is chelokot's (Sources §5).

Thirty-two squares. The same kind of cover at side 6: 13,085 weighted points of total 31.7135, so s(32) = 6. As far as we know, the first exact value of s(k² − 4) for any k ≥ 4. Kernel-checked in Lean with no hypothesis.

Twenty-one and forty-five squares. At sides 5 and 7, points alone always left a hole near some slightly tilted square. Spreading part of the weight evenly along the grid lines closes those holes: s(21) = 5 and s(45) = 7, each checked by two separately written exact checkers. s(21) has a Lean theorem from the checkers' covering statement; s(45) has no Lean yet.

Sixty squares. The same line-density cover one size up again, at side 8: 23,744 weighted points plus mass on the grid lines, total 59.8587, so s(60) = 8, and since a packing of 61 contains one of 60, s(61) = 8. Checked by the same two exact checkers; no Lean yet.

Three short of any square. Instead of a new cover for each box, one pattern that stretches: fixed corner pieces, a band along each wall that repeats with period 1, and plain area inside, all on a grid of pitch 1/5. Each box of side k gets total mass k² − 3.389, and a unit square can always be shifted into the 7 × 7 box without noticing, so checking that one box proves s(k² − 3) = k for every k ≥ 6. Bentz had proved it up to k = 7 and suggested it holds for all k. The 7 × 7 check is done at margin zero by an exact program, reviewed, and independently re-checked by wand125's separately written checker; in Lean the step from that box to every k, the use of the box's symmetry and the axis-parallel squares are kernel-checked, and since October 2026 the tilted squares too: wand125's kernel checker proves the whole theorem with no hypothesis (its 9,800 generated pieces were built by wand125, not rebuilt here).

Four short of any square. The same kind of pattern with a band of width 3: total k² − 4.295 in every box, and every unit square shifts into the 9 × 9 box, so one check there proves s(k² − 4) = k for every k ≥ 8; with s(21), s(32) and s(45) above, for every k ≥ 5, which is the deficit-4 case of Friedman's Conjecture 1. No exact value of s(k² − 4) for k ≥ 4 was known before September 2026; wand125 proved s(77) = 9 first, on 1 October. The 9 × 9 check is done at margin zero by the same program as for k² − 3, with three adversarial reviews; its tilted part has since been re-checked by wand125's separately written exact checker (6 Oct), reviewed by jlevy/squares but not yet register-verified. Lean covers everything else, as for k² − 3.

Short of a square, at scale. How many squares can be removed from a k × k grid while side k is still needed? Upper bounds on that number, c*(k), are packings of k² − c squares in a smaller square, for every k at once. Checked in Lean with no hypothesis: Kevin Fang's s(n² − n + 1) < n (n ≥ 15), Ryu's c*(k) ≤ 8⌈√(k − 4)⌉ − 1 and his k2/5 bound (for k ≥ 8·10⁶), and two new explicit bounds from Ryu's L-shaped packing with filled band ends: c*(k) ≤ 4⌈√(2k − 7)⌉ − 1 for every k ≥ 7 and c*(k) < 18.9 k2/5 + 35.7 k1/5 + 122 for every k ≥ 3·10⁵.

Squares in rectangles. Which rectangles hold three, four or five unit squares, at any angles? The boundaries, with tilted pieces from three squares on, are constructions in Erich Friedman's Math Magic problem of June 2010. That nothing below them can be packed is proved here, in Lean with no hypothesis (October 2026): in strips lower than 2 by a chain argument, each square held back by shelves that its neighbours make; with both sides at least 2 by weighted points (for five squares, ten points of weight 2/3 and 1/3, plus the four points of the classical proof of s(5) = 2 + 1/√2); and one step for five squares by generated polynomial certificates.

Exact forms of the record packings. Upper bounds, not floors: for each of the 324 records on jlevy/squares' register (n ≤ 324, as of 6 October), we solved the packing's contact graph exactly. For 269 of them the side is now given exactly, as a root of an explicit irreducible integer polynomial in a rational interval, realised by an exact configuration that an independent checker verifies. All 324 have a certified rational upper bound. Lean's kernel checks that the exact configuration is a packing (Packs n S*) for 258, and that the packing is a local minimum of the side for 178: the 176 records of integer side by one lemma, and s(11) and s(28) separately (7 October 2026; jlevy/squares#419, #420). Several records have been beaten since (jlevy/squares#401, and ours, #399); for those n the exact form describes a packing that is no longer the best.

How they are checked. Each result is a statement about a continuum of positions and angles, so the checking is the proof. The certificates are exact files of integers. Each is checked by two programs that share no code (they do share the formulation of the point test); where the margin is zero they subdivide the space of placements until every piece is proved, with nothing sampled. Lean's kernel re-checks the covering property itself for s(11) ≥ 3040/797, s(12) ≥ 35/9, ≥ 3920/997 and ≥ 15680/3951, s(13) = 4 and s(32) = 6; for s(21) and the k² − 3 and k² − 4 families Lean proves the result from the cover's covering statement (for the families, from its tilted part only), for s(45) and s(60) there is no Lean yet, and the cover rests on the checkers. The two families are the exception to "two programs": each box is checked by one program, the same for both (six adversarial reviews of the code, three more of the k² − 4 run); the k² − 3 box has since been re-checked by wand125's separately written checker. None of this is peer reviewed.

Others have used weighted certificates in 2026 too: on s(17) (Burns, Massaccesi, Mira, Guzhou0806, Kleddamag), on s(11) (jlevy, Kleddamag, Queuingtheorydotcom, who settled it), and with rectangle densities for many n from 18 to 95 (tokoharu, wand125). Sources §5 lists them.